I Tested Claude vs ChatGPT vs Gemini on Legal Drafting: Winner Shocked Me
Same Contract. Three AIs. One Catastrophic Difference in the Indemnification Clause.
I gave Claude, ChatGPT, and Gemini the exact same prompt to draft a standard freelance service agreement. Same instructions, same context, same contract structure. Two of them produced documents that looked professional — clean formatting, reasonable language, nothing obviously wrong. But when I got to the indemnification clause, one of them quietly included language that could expose the freelancer to unlimited liability for things entirely outside their control. If you'd signed that contract without a lawyer catching it, you'd have had no idea. This article breaks down exactly what each AI got right, what it missed, and — more importantly — how to prompt any of these tools so you get legally tighter output from the start. If you're using AI to draft, review, or understand contracts right now, this matters more than you think.
How I Set Up the Test (and Why the Prompt Wording Was Everything)
Every comparison test lives or dies by whether the conditions are actually equal. I used the same base prompt across all three tools, word for word: "Draft a freelance service agreement for a web designer providing a one-time website build for a small business client. Include payment terms, IP ownership, revision limits, and an indemnification clause. The contract should favor the freelancer."
That last sentence — "favor the freelancer" — was the critical instruction. It's the kind of qualifier a real person would include. It also immediately separated the AIs by how well they could apply a legal perspective, not just legal language.
ChatGPT (GPT-4o) produced a confident, well-structured contract in about 15 seconds. The formatting was excellent. Payment terms were clear. The IP clause correctly assigned ownership to the client only upon full payment — a freelancer-friendly standard move. But the indemnification clause was broad and bilateral, meaning both parties indemnified each other for almost everything. On the surface, that sounds fair. In practice, it leaves the freelancer exposed if the client claims the website caused them business losses — even losses the designer had nothing to do with.
Gemini (1.5 Pro) generated a shorter document and hedged constantly. Every third clause included a phrase like "consult a licensed attorney in your jurisdiction." That's technically responsible, but it made the output feel like a rough draft with footnotes rather than a usable document. The indemnification clause it produced was vague — it didn't clearly cap the freelancer's liability or carve out third-party IP claims, which is one of the most important protections a web designer needs.
Claude (Sonnet 3.5) got the indemnification clause right. It included a mutual indemnification structure with a carve-out for third-party IP infringement claims initiated by the client's own content — exactly the kind of specific protection a freelance web designer needs. When the client gives you their logo, their copy, and their images, and then someone sues over copyright, you don't want to be holding the bag. Claude addressed that without me asking for it specifically.
The Hidden Skill You're Not Testing: Legal Perspective, Not Just Legal Language
Most people test AI legal tools by asking "does it sound like a contract?" That's the wrong question. The real question is: does it apply a coherent legal perspective to protect the right party?
Any of these three tools can reproduce standard contract boilerplate. The internet is full of it, and all three models were trained on it. What separates them is whether they can reason about which clauses matter most given your specific situation — and then write those clauses accordingly.
Think of it this way: a contract is a series of decisions about who bears risk. IP ownership, payment triggers, liability caps, indemnification scope — each clause is allocating risk between two parties. If you tell an AI "favor the freelancer," it needs to understand which risks to shift away from your client and toward the other side. That requires more than pattern-matching on legal templates.
Claude consistently performed best at this kind of directional reasoning. When I followed up with "Rewrite the indemnification clause so the freelancer is only liable for their own direct negligence, not consequential damages or third-party claims," Claude produced clean, specific language immediately. ChatGPT produced something similar but broader. Gemini's rewrite was almost identical to its original — it didn't seem to process the distinction between direct and consequential damages clearly.
This gap matters most in clauses that deal with liability caps, indemnification triggers, and IP warranties — the three places where a contract can hurt you years after you sign it.
How to Get Better Legal Output From Any of These Tools Starting Today
You don't have to wait for a perfect AI to get better contract drafts. The gap between "good output" and "dangerous output" usually comes down to how you prompt.
Step 1: State the perspective explicitly. Don't just say "draft a contract." Say "draft a contract that strongly favors the service provider and minimizes their exposure to consequential damages." Perspective language forces the AI to make directional choices instead of splitting every clause down the middle.
Step 2: Ask for a clause-by-clause risk breakdown. After the AI drafts the contract, send this follow-up: "Review each clause in this contract and flag any language that could create unexpected liability for the freelancer. Explain the risk in plain language." This works extremely well in Claude — it'll identify its own blind spots and explain them clearly.
Step 3: Use the "devil's advocate" prompt. This one is underrated. After you have a draft, try: "You're a lawyer representing the client. Which clauses in this contract would you push back on, and what would you change to favor your client instead?" This flips the perspective and exposes weak points in your draft before the other party's real lawyer does.
Step 4: Always specify jurisdiction. Contract law varies significantly by state and country. Adding "This contract will be governed by the laws of California" (or wherever applies) changes the output noticeably — especially for enforceability language around non-competes and liability waivers.
Run through these four steps on any contract draft and you'll produce something meaningfully tighter than what you'd get from a single generic prompt. You're not replacing a lawyer — but you're arriving at that lawyer's desk with a much stronger starting document.
The Part Most People Get Wrong
Most people use AI for legal drafting by typing something like "write me a freelance contract" and then lightly editing whatever comes out. That's the mistake. You're treating the AI like a document generator when you should be treating it like a first-draft lawyer you can interrogate.
The output from a generic prompt isn't wrong because the AI is bad — it's wrong because a vague prompt produces averaged-out, risk-neutral language. Averaged-out language protects no one specifically, which means it protects you poorly.
The second big mistake is trusting formatting over substance. A well-formatted contract with clean section headers and professional language feels solid. All three AIs produced contracts that looked like real legal documents. But the indemnification clause in ChatGPT's draft had language that could be read as making the freelancer liable for the client's lost profits — a consequence that has nothing to do with how the document looked.
Read every clause out loud and ask: "Who bears the risk here?" If the answer is unclear, or if it's you when it shouldn't be, send it back to the AI with a specific instruction to fix that clause. One targeted follow-up prompt is worth more than ten generic drafts.
Key Takeaways
- Claude won the indemnification test: It was the only tool that proactively included a third-party IP carve-out without being asked — the clause most likely to protect a freelancer in a real dispute.
- ChatGPT excels at structure and speed: It produces the most polished-looking contracts fastest, but its default risk allocation leans neutral, not favorable to either party.
- Gemini hedges too much to be immediately useful: Its constant legal disclaimers make it better as a research assistant than a drafting tool for contracts you actually plan to use.
- Perspective prompting changes everything: Adding "favor the [party]" and "minimize exposure to consequential damages" to your prompt produces dramatically tighter, more protective language from any tool.
- The devil's advocate follow-up is your best quality check: Asking the AI to argue against its own contract from the other party's perspective reveals weaknesses faster than re-reading the document yourself.
What to Do Right Now
Open Claude right now and paste this prompt: "Draft a freelance service agreement for [your service] favoring the service provider. Include a mutual indemnification clause that carves out third-party IP claims arising from client-provided content, limits the freelancer's liability to the total contract value, and assigns IP to the client only upon receipt of full payment. Governing law: [your state]." Then send the follow-up: "Flag any clauses that could create unexpected liability for the service provider and suggest tighter language." You'll have a genuinely usable first draft in under five minutes — one that's starting from a much stronger position than anything a generic prompt would produce.